torch.pow - tensor(0, device='xla:1')
toch.float_power - tensor(0.6058, device='xla:1')
** - 0.6057704364907279

HloModule SyncTensorsGraph.5, entry_computation_layout={(s64[],s64[])->(s64[])}

ENTRY %SyncTensorsGraph.5 (p0: s64[], p1: s64[]) -> (s64[]) {
  %p1 = s64[] parameter(1), frontend_attributes={neff_input_name="input1"}
  %p0 = s64[] parameter(0), frontend_attributes={neff_input_name="input0"}
  %power = s64[] power(s64[] %p1, s64[] %p0)
  ROOT %tuple = (s64[]) tuple(s64[] %power), frontend_attributes={neff_output_names="output0"}
}

HloModule SyncTensorsGraph.6, entry_computation_layout={(s64[],f32[])->(f32[])}

ENTRY %SyncTensorsGraph.6 (p0: s64[], p1: f32[]) -> (f32[]) {
  %p1 = f32[] parameter(1), frontend_attributes={neff_input_name="input1"}
  %p0 = s64[] parameter(0), frontend_attributes={neff_input_name="input0"}
  %convert = f32[] convert(s64[] %p0)
  %power = f32[] power(f32[] %p1, f32[] %convert)
  ROOT %tuple = (f32[]) tuple(f32[] %power), frontend_attributes={neff_output_names="output0"}
}

import torch
import torch_xla.core.xla_model as xm

beta = 0.995
def func1(step):
    step = torch.tensor(step, device=xm.xla_device())
    factor = torch.pow(beta, step)
    print(factor)

def func2(step):
    step = torch.tensor(step, device=xm.xla_device())
    factor = torch.float_power(beta, step)
    factor = factor.float()
    print(factor)

def func3(step):
    factor = (beta ** step)
    print(factor)

func1(100)
func2(100)
func3(100)